Enable simplify qdq to work with FP8 types (#2528)

src/include/migraphx/op/dequantizelinear.hpp

@@ -72,8 +72,8 @@ struct dequantizelinear
         visit_all(x, x_zero_point)([&](auto input, auto zero_pts) {
             visit_all(result, x_scale)([&](auto output, auto scales) {
                 par_for(output_shape.elements(), [&](auto i) {
-                    output[i] = static_cast<double>(static_cast<int64_t>(input[i]) -
-                                                    static_cast<int64_t>(zero_pts[i])) *
+                    output[i] = static_cast<double>(static_cast<double>(input[i]) -
+                                                    static_cast<double>(zero_pts[i])) *
                                 scales[i];
                 });
             });

src/include/migraphx/op/quantizelinear.hpp

@@ -80,10 +80,10 @@ struct quantizelinear
                 auto min_value   = std::numeric_limits<quant_type>::min();
                 auto max_value   = std::numeric_limits<quant_type>::max();
                 par_for(output_shape.elements(), [&](auto i) {
-                    int64_t quantized = static_cast<int64_t>(std::nearbyint(input[i] / scales[i])) +
-                                        static_cast<int64_t>(zero_pts[i]);
-                    output[i] = std::max(static_cast<int64_t>(min_value),
-                                         std::min(static_cast<int64_t>(max_value), quantized));
+                    double quantized = static_cast<double>(std::nearbyint(input[i] / scales[i])) +
+                                       static_cast<double>(zero_pts[i]);
+                    output[i] = std::max(static_cast<double>(min_value),
+                                         std::min(static_cast<double>(max_value), quantized));
                 });
             });
         });

src/rewrite_quantization.cpp

@@ -58,8 +58,8 @@ void apply_quantizelinear(module& m, instruction_ref ins)
         add_zero_point = m.insert_instruction(ins, make_op("add"), add_zero_point, zero_point);
     }
 
-    int64_t max_quant = 0;
-    int64_t min_quant = 0;
+    double max_quant = 0;
+    double min_quant = 0;
     ins->get_shape().visit_type([&](auto qt) {
         max_quant = qt.max();
         min_quant = qt.min();
@@ -70,8 +70,8 @@ void apply_quantizelinear(module& m, instruction_ref ins)
 
     if(enabled(MIGRAPHX_ENABLE_CK_WORKAROUNDS{}))
     {
-        std::vector<int> min_data(s.elements(), min_quant);
-        std::vector<int> max_data(s.elements(), max_quant);
+        std::vector<double> min_data(s.elements(), min_quant);
+        std::vector<double> max_data(s.elements(), max_quant);
         min_arg = m.add_literal(literal(s, min_data));
         max_arg = m.add_literal(literal(s, max_data));
     }

src/simplify_qdq.cpp

@@ -82,18 +82,21 @@ struct match_find_quantizable_ops
     // Helper function to insert quantized versions of any broadcasts and transpose ops that
     // occur between dequantizelinear and the quantized op
     static auto
-    propagate_quantized_ins(module& m, const instruction_ref dqins, const instruction_ref qop)
+    propagate_quantized_ins(module& m, const instruction_ref dqins, const instruction_ref qop_arg)
     {
-        auto qinp     = dqins->inputs().front();
-        auto next_ins = dqins;
-
-        while(next_ins != qop)
+        auto prev_ins = qop_arg;
+        std::vector<instruction_ref> ins_inbetween;
+        // matcher skips continguous, multi/broadcasts and transposes, collect all those
+        // instructions
+        while(prev_ins != dqins)
         {
-            if(next_ins->name() != "dequantizelinear")
-            {
-                qinp = m.insert_instruction(qop, next_ins->get_operator(), qinp);
-            }
-            next_ins = next_ins->outputs().front();
+            ins_inbetween.push_back(prev_ins);
+            prev_ins = prev_ins->inputs().front();
+        }
+        auto qinp = dqins->inputs().front();
+        for(auto ins : reverse_iterator_for(ins_inbetween))
+        {
+            qinp = m.insert_instruction(dqins, (*ins)->get_operator(), {qinp});
         }
         return qinp;
     }
@@ -124,10 +127,11 @@ struct match_find_quantizable_ops
         auto scale2 = r.instructions["scale2"];
         auto zp1    = r.instructions["zp1"];
         auto zp2    = r.instructions["zp2"];
-
-        // Only INT8 type currently supported
-        if(dq1->inputs().front()->get_shape().type() != migraphx::shape::int8_type or
-           dq2->inputs().front()->get_shape().type() != migraphx::shape::int8_type)
+        // Only INT8 or FP8 type currently supported
+        std::set<migraphx::shape::type_t> supported_types = {migraphx::shape::fp8e4m3fnuz_type,
+                                                             migraphx::shape::int8_type};
+        if(not contains(supported_types, dq1->inputs().front()->get_shape().type()) or
+           not contains(supported_types, dq2->inputs().front()->get_shape().type()))
             return;
 
         // Only symmetric quantization supported (ie. non-zero zero_points not allowed)
@@ -140,8 +144,8 @@ struct match_find_quantizable_ops
 
         // Propagate q1 and q2 through any broadcasts and transposes before qop
         auto qop_args  = qop->inputs();
-        qop_args.at(0) = propagate_quantized_ins(m, dq1, qop);
-        qop_args.at(1) = propagate_quantized_ins(m, dq2, qop);
+        qop_args.at(0) = propagate_quantized_ins(m, dq1, qop_args[0]);
+        qop_args.at(1) = propagate_quantized_ins(m, dq2, qop_args[1]);
         instruction_ref dq;
         instruction_ref out_scale;
         instruction_ref zero_point;

test/simplify_qdq_test.cpp

@@ -527,6 +527,62 @@ TEST_CASE(dot_add)
     EXPECT(m1 == m2);
 }
 
+TEST_CASE(dot_add_multiple_dq_use)
+{
+    migraphx::shape sh1{migraphx::shape::float_type, {32, 1}};
+    migraphx::shape sh2{migraphx::shape::float_type, {32, 32}};
+    migraphx::module m1;
+    {
+        auto t1    = m1.add_parameter("t1", sh1);
+        auto t2    = m1.add_parameter("t2", sh2);
+        auto scale = m1.add_literal(0.5f);
+        auto zero  = m1.add_literal(std::int8_t{0});
+
+        auto q1 = add_quantize_op(m1, "quantizelinear", t1, scale, zero);
+        auto d1 = add_quantize_op(m1, "dequantizelinear", q1, scale, zero);
+        auto d1_t =
+            m1.add_instruction(migraphx::make_op("transpose", {{"permutation", {1, 0}}}), d1);
+        auto d1_tmb =
+            m1.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {32, 32}}}), d1_t);
+        auto d1_tmbc = m1.add_instruction(migraphx::make_op("contiguous"), d1_tmb);
+        auto q2      = add_quantize_op(m1, "quantizelinear", t2, scale, zero);
+        auto d2      = add_quantize_op(m1, "dequantizelinear", q2, scale, zero);
+        auto dot_1   = m1.add_instruction(migraphx::make_op("dot"), d1_tmbc, d2);
+        auto q3      = add_quantize_op(m1, "quantizelinear", dot_1, scale, zero);
+        auto d3      = add_quantize_op(m1, "dequantizelinear", q3, scale, zero);
+        auto dot_2   = m1.add_instruction(migraphx::make_op("dot"), d3, d1);
+        auto add     = m1.add_instruction(migraphx::make_op("add"), {dot_2, d1});
+        m1.add_return({add});
+    }
+
+    migraphx::module m2;
+    {
+        auto t1    = m2.add_parameter("t1", sh1);
+        auto t2    = m2.add_parameter("t2", sh2);
+        auto scale = m2.add_literal(0.5f);
+        auto zero  = m2.add_literal(std::int8_t{0});
+
+        auto q1 = add_quantize_op(m2, "quantizelinear", t1, scale, zero);
+        auto q1_t =
+            m2.add_instruction(migraphx::make_op("transpose", {{"permutation", {1, 0}}}), q1);
+        auto q1_tmb =
+            m2.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {32, 32}}}), q1_t);
+        auto q1_tmbc     = m2.add_instruction(migraphx::make_op("contiguous"), q1_tmb);
+        auto q2          = add_quantize_op(m2, "quantizelinear", t2, scale, zero);
+        auto dot_1       = m2.add_instruction(migraphx::make_op("quant_dot"), q1_tmbc, q2);
+        auto out_scale   = add_scale_mul(m2, scale, scale, 1, 1, dot_1->get_shape().lens());
+        auto d3          = add_quantize_op(m2, "dequantizelinear", dot_1, out_scale);
+        auto d3_q        = add_quantize_op(m2, "quantizelinear", d3, scale, zero);
+        auto dot_2       = m2.add_instruction(migraphx::make_op("quant_dot"), d3_q, q1);
+        auto out_scale_2 = add_scale_mul(m2, scale, scale, 1, 1, dot_2->get_shape().lens());
+        auto d4          = add_quantize_op(m2, "dequantizelinear", dot_2, out_scale_2);
+        auto add         = m2.add_instruction(migraphx::make_op("add"), d4, t1);
+        m2.add_return({add});
+    }
+    run_pass(m1);
+    EXPECT(m1 == m2);
+}
+
 TEST_CASE(conv)
 {
     migraphx::shape s4{migraphx::shape::int8_type, {1280, 320, 1, 1}};
@@ -919,7 +975,6 @@ TEST_CASE(mobilenet_snippet)
 
     auto mod1 = create_module();
     auto mod2 = create_module();
-
     run_pass(mod2);
 
     auto match_qdq = migraphx::match::name("dequantizelinear")(